Research On Joint Algorithm Of Intention Recognition And Slot Filling In Task-based Conversation

Nowadays,the development of social science and technology is obvious,and deep learning has been applied in natural language processing.The research of dialogue system is becoming more and more important,attracting a lot of people's attention and widely used in people's life,especially task-based dialogue system for specific fields can help companies greatly save costs.Natural language understanding,state tracking and natural language generation are three main parts of the dialogue system.The natural language semantic understanding module is to extract information from the input scripts,so that the machine can better carry out the follow-up tasks.It needs to transform the input unstructured information into the structured information that the machine can understand.Intention recognition and slot filling are very important and basic tasks in the module.In the current model,the two tasks are generally modeled separately,ignoring the correlation between them,and the intention recognition and slot filling tasks are not well combined.Combining the two tasks can not only reduce the use of parameters,but also improve the performance.Therefore,this paper puts forward the corresponding model around how to strengthen the joint between the two.This paper studies the joint model of intention recognition and slot filling from two aspects:First,through research and experiments,we designed and implemented a variety of baseline algorithms for intention recognition and slot filling tasks,and built a task-based dialogue system;at the same time,we also implemented a joint algorithm model based on parameter sharing,and analyzed and discussed the existing problems of the current joint model.Second,based on the pre training mechanism,the fusion of intention recognition and slot filling model is realized in the bottom coding layer.In this paper,we use Bert model to implement the joint model,and use Bert to replace the bottom word vector representation model in intention recognition and slot filling baseline model.The bottom parameters of the two models are shared by Bert,which enhances the feature representation of the input text.The experimental results show thatour model have good results on the data set.Third,based on the guiding fusion mechanism,the model fusion is realized at the output of intention recognition and slot filling.The encoded intention recognition results are added to the slot information representation vector of each word through linear transformation,and the intention information is used to guide the slot filling module,so as to strengthen the combination of intention recognition and slot filling.Tests based on ATIS and Snips data show that the model has achieved good performance.Fourth,according to the above innovative joint model,a Chinese dialogue system for printer after-sales repair is built.The system is to relieve the customer service pressure,and can answer the user's questions better and quickly.The system uses the latest innovation model,which can complete the Q&A interaction with users well.